遥感
计算机科学
分割
人工智能
计算机视觉
特征(语言学)
图像分割
遥感应用
特征提取
高光谱成像
光谱特征
模式识别(心理学)
地质学
合成孔径雷达
像素
光谱分析
迭代重建
反射率
图像处理
作者
Yufei Gao,Zhuangzhi Nie,Liuqian Wang,Yucheng Shi,Dan Zhang,Qi Ming
标识
DOI:10.1109/tgrs.2026.3741907
摘要
High-resolution remote sensing (HRRS) image segmentation requires broad semantic context together with accurate boundaries and small-object geometry. H-CoReNet combines three complementary operations for this purpose. The spectral frequency oscillator (SFO) summarizes the magnitude of a 2-D real Fourier transform and uses the resulting descriptor to gate a parallel spatial branch. The harmonic coherence unit (HCU) dynamically weights$3\times 3$,$5\times 5$, and$7\times 7$depthwise responses and reconstructs the bottleneck feature from an adaptive shared dictionary. The Long-Distance Roaming Window Association (LDRWA) integrates deformable window sampling, mean-centered local modulation, and learnable complex spectral filtering. Experiments use UAVid, an uncrewed aerial vehicle (UAV) urban-scene semantic segmentation benchmark, together with Potsdam, Vaihingen, and LoveDA. Matched test-time augmentation (TTA) and single-scale inference without augmentation (no-TTA) protocols are reported with region- and boundary-sensitive metrics. The quantitative comparison includes recent frequency-domain models, including AFENet and CFormer. Under the matched$512~\times ~512$UAVid profiling protocol, the ConvNeXt-Base configuration contains 89.7M parameters and 100.6G FLOPs; the three proposed modules add 0.4M parameters to the backbone-only configuration. The results show that magnitude-guided gating, adaptive dictionary reconstruction, and deformable spectral filtering provide complementary information for remote sensing segmentation. Code is available athttps://github.com/ZhuangzhiNie/H-CoReNet
科研通智能强力驱动
Strongly Powered by AbleSci AI